For self-hosted, LiveKit documents this:
Adaptive interruption uses LiveKit Cloud inference.
Self-hosted usage includes 40,000 inference requests/month.
For higher volume, they ask you to contact sales.
What happens exactly after quota is exhausted is not currently explicitly documented as a hard error vs silent fallback. Can someone tell what happens when quota is reached? Does it switch to normal VAD based?
If an agent exceeds the included limit, it will no longer have adaptive interruption handling. The agent will gracefully fall back to non-adaptive, which is VAD.
Good call out on the doc gap. We will add clarification to the docs.
Is there an option to pay for additional inference requests for self agents?
No, we do not currently provide that as an option.
Hello @CWilson , the VAD is not performing as good as Adaptive and we are currently on a Paid plan. Can we pay for additional requests for self hosted deployment for Adaptive interruption? Let me know because we want best experience for our customers and currently due to VAD its not a good experience.
@nilkanth I’m afraid not - adaptive interrupt handling is a cloud-only feature.
May I ask what is stopping you moving from self-hosted agents to LiveKit Cloud? We want LiveKit Cloud to be the best experience, so are always keen to understand what is blocking developers from moving.
Most of our clients are in heavily regulated industries like banking, insurance, health. Hence they prefer self hosted / on-prem solutions.
@nilkanth Thanks for the feedback and that makes total sense. We are continuing to work on features that will make cloud-hosting more amenable to regulated industries but there is still work to do.
the VAD is not performing as good as Adaptive
Just to clarify, you can still use the turn detector model with self-hosted agents, LiveKit turn detector plugin | LiveKit Documentation, which is better than VAD, it’s just the adaptive interruption handling that is cloud-only.